Research map / Image generation / Text-to-image diffusion

EasyCraft: A Robust and Efficient Framework for Automatic Avatar Crafting

CVPR 2025 · Image generation › Text-to-image diffusion · approach: Multi-subject personalization · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes EasyCraft, a feedforward translator (MAE-pretrained ViT encoder plus parameter MLPs) that maps photos of any style to game avatar crafting parameters, and pairs it with an engine-style fine-tuned Stable Diffusion for text-to-avatar creation.

Problem

Automatic game-engine avatar (face) crafting from photos or text is hard because prior unsupervised methods depend on engine-specific image-domain constraints and generalize poorly across input styles and engines.

Approach

Self-supervised MAE pretraining of a ViT encoder on a multi-style face dataset gives a unified feature distribution, so a translator trained only on engine-rendered parameter-screenshot pairs (encoder frozen) transfers to photos of any style; text input is handled by fine-tuning SD on engine-style captioned renders and feeding its outputs to the translator.

Key results

• Photo-based: FID 17.65 on Justice Mobile vs 40.69 (F2P) and 34.27 (F2P v2), with 0.026s inference per photo on RTX 4090 versus 1.140s for F2P.
• Text-based: LPIPS 0.093 vs 0.027 for T2P on Justice Mobile, and FID 18.76 vs 32.9, with 0.643s vs 1.725s speed.
• User study: 87% preference for EasyCraft in photo-based comparison and 77% in text-based comparison.

Measured effects reported in the paper

Surprising findings

• Removing ViT pretraining raised the inception score (2.060 vs 1.216) but sharply worsened identity similarity and FID, because it produced many distorted avatars.
• ViT pretraining had almost no effect on text-based creation, since the engine-style SD already produces images in the engine style.

Limitations

• Performance on identity similarity is slightly below F2P (0.351 vs 0.376 on Justice Mobile), which uses direct identity supervision.
• Evaluated only on two mobile RPG engines; generalization to other engines is claimed but not demonstrated.
• Text-based pipeline depends on GPT-4o captions and a small 7,000-image fine-tuning set.

Experimental setup

Datasets: AffectNet CASIA-WebFace CelebA IMDB-WIKI SeePrettyFace AnimeFace Danbooru Justice Mobile engine renders Naraka: Bladepoint Mobile engine renders

Benchmarks: Justice Mobile test photos (1,000 images) Naraka: Bladepoint Mobile test photos text prompt set (200 descriptions)

Metrics: identity similarity (ArcFace) inception score FID LPIPS CLIP score inference speed user study fidelity and preference

Baselines: F2P F2P v2 T2P

Builds on: Masked Autoencoders Stable Diffusion

Closest papers by meaning

• Unbounded: A Generative Infinite Game of Character Life Simulation (ICLR 2025 · similarity 0.83)
• PERSE: Personalized 3D Generative Avatars from A Single Portrait (CVPR 2025 · similarity 0.82)
• ShapeCraft: LLM Agents for Structured, Textured and Interactive 3D Modeling (NeurIPS 2025 · similarity 0.82)
• ChatGarment: Garment Estimation, Generation and Editing via Large Language Models (CVPR 2025 · similarity 0.82)
• CharacterBench: Benchmarking Character Customization of Large Language Models (AAAI 2025 · similarity 0.82)
• MineObserver 2.0: A Deep Learning & In-Game Framework for Assessing Natural Language Descriptions of Minecraft Imagery (AAAI 2024 · similarity 0.82)